Digital technology: Options for post-market surveillance

Real-world data from electronic health records (EHRs), registries, and wearables are promising sources for post-market surveillance. Regulators are cautiously receptive – what is needed to reassure them? Real-world data from electronic health records (EHRs), registries, and wearables are promising sources for post-market surveillance. Regulators are cautiously receptive – what is needed to reassure them?
The UK National Institute for Health and Care Excellence (NICE) bridges randomised and real-world data in its Cancer Drugs Fund. Companies collect real-world data for NICE on clinically effective drugs that are not cost-effective at currently accepted thresholds. An analogous programme is running for digital psychological therapies. NICE’s statement of intent for data analytics sets out when and why broader types of data should be considered. Challenges include how to recognise a high-quality analysis and understanding when real-world data is an acceptable alternative. Companies want to know what evidence NICE will accept without requiring a formal trial.

In March, NICE published its evidence standards framework for digital health technologies, outlining tiers of evidence required according to the potential risk of an intervention. For high-risk items, best practice is a randomised controlled trial (RCT). A pilot assessing four digital technologies is underway, including one to detect cardiac arrhythmias. All use the Cancer Drugs Fund model, whereby an initial core level of evidence is required, followed by real-world data collection by the company, and re-evaluation by NICE.

From an industry perspective, post-market surveillance should harness the options digital health offers. For example, fully digital spontaneous adverse event reporting along with social media to identify events early; wearables and internet-of-things for real-time monitoring; and routine use of registries and EHRs. The Innovative Medicines Initiative (IMI) WEB-RADR project states that social media channels may be a useful adjunct to pharmacovigilance in niche areas, and adverse event recognition algorithms could broaden the scope. Mobile apps for patients to directly report drug side effects are only used if they are simple and intuitive to use and provide Information to patients – e.g. drug safety alerts.

Companies are less willing to invest in patient-facing technologies for data collection as long as there is no clarity from regulators on its acceptability. Regulator buy-in is also needed to exploit the advantages of registry-based RCTs – they are positive in principle, but the onus appears to be on companies to give examples. One way to provide proof of quality is a hybrid model of traditional adverse event reporting coupled with registry data, followed by comparison and publication.

When it comes to processing and analysing data, artificial Intelligence (AI) has the potential to detect safety signals in complex datasets, while bots and rules-based automation could replace manual tasks in case handling. Digitalisation and automation improve quality and reduce costs and can also be used to send information to regulators, ethics committees, healthcare professionals, and patients.

The Carelink network is a cloud-based platform collecting remote monitoring data from patients with implantable devices. It provides insights into on-label and off-label use and enables estimation of underreporting from complaint handling. This could be linked with other platforms holding outcome data – for example Hugo, in which patients authorise sharing of their health data – to identify associations between a product and outcomes. Developments in post-market surveillance need to be accompanied by efforts to guarantee data security, including encryption and privacy.

Discussion points

• Regulators sound enthusiastic about new methods of collecting information until it comes to licensing or marketing authorisation and then they become precedent- rather than novelty-related. Where are we going? Companies believe it’s up to them to provide examples.
• Regulators are confident about RCTs because they know what a good quality trial looks like and bias is eliminated. With other data, they believe it is hard to ascertain quality and the impact of bias.
• In future companies may receive provisional approval based on some data with caveats until real-world data are collected.
• A potential model of market authorisation for cardiovascular medicine, where many diseases are common, is to give limited approval for a defined, small patient population. Then conduct early post-market surveillance to generate more safety data, instead of a huge, costly phase III trial. The indications for approval could be expanded later.
• Are there specifications envisaged for evaluating digital therapies? How different will that be to traditional RCTs? Experts are looking at methods to take bias out of data – some say it can be removed without an RCT; others disagree. It is likely that RCTs will not always be needed, but work is needed on when alternatives are acceptable and how to assess their quality.
• Could a hybrid of traditional data and EHRs be analysed in one country to reassure regulators that there are no untoward events, or they are predictable? The rest of the study could then proceed in other countries. A limitation is that data quality and type vary by country – Sweden and the UK have national healthcare systems with comprehensive high-quality registries; the US has claims databases.
• Serious adverse events caused by a treatment will leave a trace in the healthcare system. It means this information can be collected from routine data in connected health systems and additional monitoring is unnecessary. We need to make sure these events are recorded; collated and analysed; and communicated to the regulator.
• Social media has picked up adverse events not detected by routine data collection. We need a loop back to regulators when that occurs.
• Adverse events are passively collected in a clinical dataset – meaning that if it is reported, it happened; if it is not reported, it effectively did not happen. Prospective studies like RCTs actively ask participants: did you have this event, yes or no?
• Registry sweeps and claims databases provide more accurate adverse event data than traditional follow-up questionnaires since patients may forget events. But regulators feel more comfortable with traditional forms of data because they know how to handle it. Companies need to do both in parallel then build regulators’ confidence level.
• Attribution of adverse events to an implanted device is difficult using routine data, especially in patients who have multiple procedures during one hospitalisation, all of which could cause that event. This problem also occurs with traditional reporting. One solution is to have an adjudicator assess registry events in the same way as a clinician reports routine data.
• AI can be used to look at attributable effects in observational data. With large datasets, an infinite number of possible pathways can be explored.

Conclusion

Regulators would like faster data collection to allow them to accelerate their own assessments. Real-world data may be the answer but there are questions around quality and bias. The future may see provisional approvals followed by post-market surveillance using real-world data. High-quality real-world data is needed globally, and methods are required to remove bias from observational data. Alongside